upstream-archive/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-07-23 11:20:53 -05:00
Code Issues Packages Projects Releases Wiki Activity

feat: support for --tensor-type-rules on generation modes (#932)

Browse Source
This commit is contained in:
Wagner Bruna
2025-11-16 06:07:32 -03:00
committed by GitHub
parent 742a7333c3
commit 199e675cc7
5 changed files with 59 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files

87
model.cpp
View File

@@ -1254,15 +1254,59 @@ std::map<ggml_type, uint32_t> ModelLoader::get_vae_wtype_stat() {
return wtype_stat;
}
void ModelLoader::set_wtype_override(ggml_type wtype, std::string prefix) {
static std::vector<std::pair<std::string, ggml_type>> parse_tensor_type_rules(const std::string& tensor_type_rules) {
std::vector<std::pair<std::string, ggml_type>> result;
for (const auto& item : split_string(tensor_type_rules, ',')) {
if (item.size() == 0)
continue;
std::string::size_type pos = item.find('=');
if (pos == std::string::npos) {
LOG_WARN("ignoring invalid quant override \"%s\"", item.c_str());
continue;
}
std::string tensor_pattern = item.substr(0, pos);
std::string type_name = item.substr(pos + 1);
ggml_type tensor_type = GGML_TYPE_COUNT;
if (type_name == "f32") {
tensor_type = GGML_TYPE_F32;
} else {
for (size_t i = 0; i < GGML_TYPE_COUNT; i++) {
auto trait = ggml_get_type_traits((ggml_type)i);
if (trait->to_float && trait->type_size && type_name == trait->type_name) {
tensor_type = (ggml_type)i;
}
}
}
if (tensor_type != GGML_TYPE_COUNT) {
result.emplace_back(tensor_pattern, tensor_type);
} else {
LOG_WARN("ignoring invalid quant override \"%s\"", item.c_str());
}
}
return result;
}
void ModelLoader::set_wtype_override(ggml_type wtype, std::string tensor_type_rules) {
auto map_rules = parse_tensor_type_rules(tensor_type_rules);
for (auto& [name, tensor_storage] : tensor_storage_map) {
if (!starts_with(name, prefix)) {
ggml_type dst_type = wtype;
for (const auto& tensor_type_rule : map_rules) {
std::regex pattern(tensor_type_rule.first);
if (std::regex_search(name, pattern)) {
dst_type = tensor_type_rule.second;
break;
}
}
if (dst_type == GGML_TYPE_COUNT) {
continue;
}
if (!tensor_should_be_converted(tensor_storage, wtype)) {
if (!tensor_should_be_converted(tensor_storage, dst_type)) {
continue;
}
tensor_storage.expected_type = wtype;
tensor_storage.expected_type = dst_type;
}
}
@@ -1603,41 +1647,6 @@ bool ModelLoader::load_tensors(std::map<std::string, struct ggml_tensor*>& tenso
return true;
}
std::vector<std::pair<std::string, ggml_type>> parse_tensor_type_rules(const std::string& tensor_type_rules) {
std::vector<std::pair<std::string, ggml_type>> result;
for (const auto& item : split_string(tensor_type_rules, ',')) {
if (item.size() == 0)
continue;
std::string::size_type pos = item.find('=');
if (pos == std::string::npos) {
LOG_WARN("ignoring invalid quant override \"%s\"", item.c_str());
continue;
}
std::string tensor_pattern = item.substr(0, pos);
std::string type_name = item.substr(pos + 1);
ggml_type tensor_type = GGML_TYPE_COUNT;
if (type_name == "f32") {
tensor_type = GGML_TYPE_F32;
} else {
for (size_t i = 0; i < GGML_TYPE_COUNT; i++) {
auto trait = ggml_get_type_traits((ggml_type)i);
if (trait->to_float && trait->type_size && type_name == trait->type_name) {
tensor_type = (ggml_type)i;
}
}
}
if (tensor_type != GGML_TYPE_COUNT) {
result.emplace_back(tensor_pattern, tensor_type);
} else {
LOG_WARN("ignoring invalid quant override \"%s\"", item.c_str());
}
}
return result;
}
bool ModelLoader::tensor_should_be_converted(const TensorStorage& tensor_storage, ggml_type type) {
const std::string& name = tensor_storage.name;
if (type != GGML_TYPE_COUNT) {